[TUHS] Faster cpus at any cost

[TUHS] Faster cpus at any cost

[Steve Simon]

I know this is a dangerous game to play, but what of the future?

I am intrigued by the idea of true optical computers,
perhaps the $10M super-computer will return?

GPUs have definitely taken over in some areas - even where
SIMD would not seem a good fit. My previous employer does motion
estimation of realtime video and has moved from custom electronics
and FPGAs to using off the shelf GPUs in PCs.

-Steve

Re: [TUHS] Faster cpus at any cost

[Bakul Shah]

On Fri, 29 Jun 2018 09:51:55 +0100 "Steve Simon" <steve@quintile.net> wrote:
> I know this is a dangerous game to play, but what of the future?

[In the same spirit :-)]
The spin orbit torque (SOT) MRAM devices seem very promising.

Spin tranfer torque (STT) MRAM are both seen as persistent
memories that will work beyond the feature sizes where 
current flash based CMOS devices won't work as they rely on
retaining charge, while STT/SOT decives depend on magnetism
(the spin direction is switched with a tiny bit of electrical
energy). What is more, they don't have number of write limits
of flash memory, nor the very long write times (STT/SOT writes
are on the order of 100ps to 10ns as opposed to 1us to 1ms for
flhas and takes million times less energy to write).

STT-MRAMs are already being used in small sizes (given no
write limits & high speed, it makes a good cache layer for
SSDs).  But STT devices have a number of limits that SOT don't
have.  SOTs can be written about 10 times faster and they can
be used at even smaller feaure sizes.

The really intersting part is logic gates have been constructed
using the same technologu. Unlike tradition charge based
devices where CPU and massive memory are kept separate, here
logic and memory can be on the same chip. In fact the same
device can perform logic and retain the results and the logic
can be electrically reconfigured.  These gates can be an order
of magnitude smaller (compared to 14nm FinFET CMOS) and are
ultra energy efficient => much less heat. And massive
parallelism.

Too early to tell whether this actually pans out or scales up
to billions of gates.

And of course, if we are to believe the crowd here, this will
be an utter failure since it won't really help us C programs
faster :-)

Also, I just read this stuff; I have no insight and I may have
misconstrued everything!

A reference that may be of interest:
    https://www.nature.com/articles/s41598-017-14783-1.pdf

> I am intrigued by the idea of true optical computers,
> perhaps the $10M super-computer will return?

Optics will more likely be used for a communication layer.

Re: [TUHS] Faster cpus at any cost

[Toby Thain]

On 2018-06-29 4:51 AM, Steve Simon wrote:
> I know this is a dangerous game to play, but what of the future?
> 
> I am intrigued by the idea of true optical computers,
> perhaps the $10M super-computer will return?
> 
> GPUs have definitely taken over in some areas - even where
> SIMD would not seem a good fit. My previous employer does motion
> estimation of realtime video and has moved from custom electronics
> and FPGAs to using off the shelf GPUs in PCs.

On this topic, 10 years old, but the Kogge whitepaper on "Exascale
Computing" might be of interest:

https://www.researchgate.net/publication/242366160_ExaScale_Computing_Study_Technology_Challenges_in_Achieving_Exascale_Systems

--T

> 
> -Steve
> 
> 

Re: [TUHS] Faster cpus at any cost

[Perry E. Metzger]

On Fri, 29 Jun 2018 09:51:55 +0100 "Steve Simon" <steve@quintile.net>
wrote:
> I know this is a dangerous game to play, but what of the future?
> 
> I am intrigued by the idea of true optical computers,
> perhaps the $10M super-computer will return?

One problem with optical is feature size. Even ultraviolet 200nm light
waves are pretty big compared to the feature size we now have in the
best chips (which is under 10nm at this point, though it will soon
stall out.) If you want EM waves that are close to the size of
current features, you're in the range of X-rays and aren't going to
have an easy time manipulating them.

All that said, a different kind of gedankenexperiment:

Say you wanted to store data as densely as possible. Lets ignore how
you would manage to read it and consider things on the same order of
magnitude of "as dense as we're going to get". If you stored 1s and
0s as C12 and C13 atoms in a diamond lattice, you get about
1.75e23 bits per cc, or about 12 zettabytes per cc. (Someone should
check my math.) I think it might be hard to do more than an order of
magnitude better than that. So, that's a crazy amount of storage, but
it looks like a pretty strong limit.

> GPUs have definitely taken over in some areas - even where
> SIMD would not seem a good fit. My previous employer does motion
> estimation of realtime video and has moved from custom electronics
> and FPGAs to using off the shelf GPUs in PCs.

Not surprised to hear. They're kind of everywhere at this point,
especially in scientific computation.

Perry
-- 
Perry E. Metzger		perry@piermont.com